Re: [Patch v2] ALSA: compress: allow to leave draining state when pausing in draining

[Takashi Iwai]

On Fri, 09 Jul 2021 04:08:29 +0200,
Robert Lee wrote:
> 
> Jaroslav Kysela <perex@xxxxxxxx> 於 2021年7月8日 週四 下午10:53寫道：
> >
> > On 08. 07. 21 15:47, Robert Lee wrote:
> > > Hi Takashi,
> > >
> > > It is a little complex to describe the design in detail, but try to
> > > explain simply
> > > what issue we meet.
> > >
> > > If w/o the change,  after user resumes from the pause, our system would call
> > > snd_compr_drain() or snd_compr_partial_drain() again after it returns from
> > > previous drain (when EOF reaches). Then it will block in this drain and no one
> > > wake it up because EOF has already reached. I add this change to return from
> > > the previous drain.
> >
> > It looks like that the driver does not call snd_compr_drain_notify() so the
> > state is not updated to SETUP on EOF.
> >
> We indeed call snd_compr_drain_notify() on EOF, but after return from
> wait_for _drain there is another drain again immediately.
> Looks like the system queue some states change on user space and need
> to drain again after resume from pause.
> I suppose there is different design on user space so I add the hook to
> handle diffent usage.

Right, the previous drain-in-pause implementation was purely in the
kernel side, and user-space didn't change much; after resuming from
the pause, the driver resumes exactly to the same state before the
pause (i.e. start draining again).

The difference sounds similar like the suspend/resume scheme; some
does resume by itself to the previous state while some requires the
explicit action.

> > > Actually, I am wondering how the pause-during-drain can keep the state in
> > > DRAINING. It should have a different design. :)
> >
> > I already proposed to add a new state (because it's a new state), but the
> > conservative way was elected to avoid user space changes.

Yes, the primary concern is that the compress API uses the very same
state like PCM, and if we extend PCM state, it'll be a much larger
problem.  And, even if we change the state to compress-only, it's
still an ABI incompatibility, and it has to be carefully handled not
to break the existing application (e.g. expose the new state only when
the application is really ready to handle -- introducing a new ioctl
for state or introduce a new ioctl like SNDRV_PCM_IOCTL_USER_PVERSION
that informs the ABI version the user-space understands).


Takashi